Yes — with 3.3 GB to spare
Mistral Small 3.2 24B at Q4_K_M fits your RTX 4000 Ada entirely on the GPU at 8K context, at an estimated 16 tokens per second. Past 28K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 13.3 GB
Apache 2.0
Released Jun 2025
Vision
Apache-2.0, vision-capable, and the most 24 GB-friendly of the 2025 generalists.
The VRAM budget
weights 13.3 GB
Weights 13.3 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Free 3.3 GB of 18.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 23.4 GB | 25.2 GB | — | ~3.8 | −0.1% ppl | 6.8 GB over |
| Q6_K | 18.0 GB | 19.9 GB | — | ~8.6 | −0.4% ppl | 1.5 GB over |
| Q5_K_M | 15.6 GB | 17.4 GB | 14K | 14 | −0.8% ppl | Fits |
| Q4_K_M | 13.3 GB | 15.1 GB | 28K | 16 | −1.9% ppl | Recommended |
| Q3_K_M | 10.7 GB | 12.6 GB | 45K | 20 | −5.4% ppl | Long context |
| Q2_K | 9.2 GB | 11.1 GB | 55K | 24 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ ollama pull mistral-small:24b $ OLLAMA_CONTEXT_LENGTH=8192 \ ollama run mistral-small:24b
The default. One binary, a model registry, an OpenAI-compatible port. More on Ollama.
01Download is 13.3 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03There is room to go to 28K context on this card.